Liquid AI is revolutionizing LLMs to work on edge devices like smartphones with new ‘Hyena Edge' model
Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More
Liquid AI, the Boston-based foundation model startup spun out of the Massachusetts Institute of Technology (MIT), is seeking to move the tech industry beyond its reliance on the Transformer architecture underpinning most popular large language models (LLMs) such as OpenAI's GPT series and Google's Gemini family.
Yesterday, the company announced 'Hyena Edge,' a new convolution-based, multi-hybrid model designed for smartphones and other edge devices in advance of the International Conference on Learning Representations (ICLR) 2025.
The conference, one of the premier events for machine learning research, is taking place this year in Vienna, Austria.
Hyena Edge is engineered to outperform strong Transformer baselines on both computational efficiency and language model quality.
In real-world tests on a Samsung Galaxy S24 Ultra smartphone, the model delivered lower latency, smaller memory footprint, and better benchmark results compared to a parameter-matched Transformer++ model.
Unlike most small models designed for mobile deployment — including SmolLM2, the Phi models, and Llama 3.2 1B — Hyena Edge steps away from traditional attention-heavy designs. Instead, it strategically replaces two-thirds of grouped-query attention (GQA) operators with gated convolutions from the Hyena-Y family.
The new architecture is the result of Liquid AI's Synthesis of Tailored Architectures (STAR) framework, which uses evolutionary algorithms to automatically design model backbones and was announced back in December 2024.
STAR explores a wide range of operator compositions, rooted in the mathematical theory of linear input-varying systems, to optimize for multiple hardware-specific objectives like latency, memory usage, and quality.
To validate Hyena Edge's real-world readiness, Liquid AI ran tests directly on the Samsung Galaxy S24 Ultra smartphone.
Results show that Hyena Edge achieved up to 30% faster prefill and decode latencies compared to its Transformer++ counterpart, with speed advantages increasing at longer sequence lengths.
Prefill latencies at short sequence lengths also outpaced the Transformer baseline — a critical performance metric for responsive on-device applications.
In terms of memory, Hyena Edge consistently used less RAM during inference across all tested sequence lengths, positioning it as a strong candidate for environments with tight resource constraints.
Hyena Edge was trained on 100 billion tokens and evaluated across standard benchmarks for small language models, including Wikitext, Lambada, PiQA, HellaSwag, Winogrande, ARC-easy, and ARC-challenge.
On every benchmark, Hyena Edge either matched or exceeded the performance of the GQA-Transformer++ model, with noticeable improvements in perplexity scores on Wikitext and Lambada, and higher accuracy rates on PiQA, HellaSwag, and Winogrande.
These results suggest that the model's efficiency gains do not come at the cost of predictive quality — a common tradeoff for many edge-optimized architectures.
For those seeking a deeper dive into Hyena Edge's development process, a recent video walkthrough provides a compelling visual summary of the model's evolution.
The video highlights how key performance metrics — including prefill latency, decode latency, and memory consumption — improved over successive generations of architecture refinement.
Read More The human factor: How companies can prevent cloud disasters
It also offers a rare behind-the-scenes look at how the internal composition of Hyena Edge shifted during development. Viewers can see dynamic changes in the distribution of operator types, such as Self-Attention (SA) mechanisms, various Hyena variants, and SwiGLU layers.
These shifts offer insight into the architectural design principles that helped the model reach its current level of efficiency and accuracy.
By visualizing the trade-offs and operator dynamics over time, the video provides valuable context for understanding the architectural breakthroughs underlying Hyena Edge's performance.
Liquid AI said it plans to open-source a series of Liquid foundation models, including Hyena Edge, over the coming months. The company's goal is to build capable and efficient general-purpose AI systems that can scale from cloud datacenters down to personal edge devices.
The debut of Hyena Edge also highlights the growing potential for alternative architectures to challenge Transformers in practical settings. With mobile devices increasingly expected to run sophisticated AI workloads natively, models like Hyena Edge could set a new baseline for what edge-optimized AI can achieve.
Hyena Edge's success — both in raw performance metrics and in showcasing automated architecture design — positions Liquid AI as one of the emerging players to watch in the evolving AI model landscape.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Forbes
a day ago
- Forbes
How to Future-Proof Your Workforce for AI
Stand Together Is AI poised to automate all human work — or will it bring us more meaningful, higher-paying jobs? It depends who you ask. Tech CEOs have warned of a coming tidal wave. Anthropic CEO Dario Amodei predicted AI could eliminate half of all entry-level white-collar jobs within five years. OpenAI's Sam Altman expects artificial general intelligence to surpass humans at all economically valuable tasks. Some economists see a slower story. MIT's Daron Acemoglu estimates that AI will increase U.S. productivity by just 0.05% annually over the next decade — based on the idea that most companies won't change how they operate. Predicting exactly how AI will reshape jobs might be impossible. But if you're leading a business, your priority isn't knowing precisely who's correct — it's preparing your organization and workforce for an AI-driven future. But why is forecasting AI's impact on jobs so difficult, even for experts? Understanding the forces that make it hard to predict is also what helps leaders understand how to prepare. I think of this as AI's "Three-Body Problem" — a nod to the famously unpredictable physics phenomenon, where adding a third interacting body makes future trajectories chaotic and impossible to forecast. Similarly, three interacting factors — AI capabilities, adoption speed, and job creation potential — make AI job predictions uniquely challenging. Let's briefly explore each one. 1. What will AI be able to do? Experts fundamentally disagree about how powerful — and how fast — AI models will become. Some predict artificial superintelligence could emerge as soon as 2027, driven by self-improving systems that compound capabilities exponentially. Others believe we're nearing the limits of current scaling laws, and that even human-level intelligence may still be 15 years away. Additionally, AI's abilities are uneven — superhuman in some tasks yet subhuman in others. For example, Ethan Mollick recently demonstrated that GPT-4o could ace an MBA-level strategy case with remarkable nuance — yet failed at solving a basic riddle requiring common sense logic. It's a reminder that raw capability doesn't guarantee reliability across domains. 2. How will we choose to use AI? Just because AI can automate a job — or parts of one— doesn't mean it will. The pace at which we integrate AI into the workplace will be wildly uneven. Some sectors will move as fast as the technology itself. Others will be slowed by regulation, union rules, institutional inertia, or cultural preferences. In industries with little financial pressure to cut costs, there may simply be no urgency to change. Something I think about a lot: We still have gas station attendants pumping gas in Oregon and New Jersey — not because it's technologically necessary, but because we've chosen to protect those jobs. Predicting where and how fast adoption will happen is tough, because it's ultimately a question of human behavior. 3. What new jobs will AI create? AI will create new jobs — but how many, and what kind, remains unknown, especially if it quickly automates even the roles it helps create. Historically, new technology has been a net job creator. The internet gave rise to SEO specialists and app developers, and the smartphone boom created whole industries around mobile design and delivery. But AI is different: It could quickly take over the roles it helps create, making it harder to predict the net effect on employment. Taken together, the Three-Body Problem makes AI job forecasting impossible to pin down. AI capabilities, adoption speed, and job creation are all unpredictable — yet understanding these complexities is only half the battle. When nobody can confidently predict AI's trajectory, how do you prepare your business and your employees for what's ahead? Precisely because AI's trajectory is uncertain, organizations must build the muscle to adapt quickly. Readiness is the best hedge against uncertainty. As Bijal Shah, CEO of Guild, put it, 'Employers who win in this new world of work driven by AI will be those who invest in making their workforce more adaptable and resilient — not just for today's skills, but for tomorrow's unknowns.' So, where should leaders begin? The following strategies are no-regrets investments — moves that deliver value no matter how fast AI change arrives. They offer ways not just to survive the shifts ahead, but to build more agile organizations and more meaningful, purposeful work for employees. Forget titles. Jobs are a collection of skills and abilities. Titles make jobs sound fixed. But they're really collections of skills and abilities — Lego pieces —that can be reassembled creatively. Organizations like Censia are already shepherding adaptations — such as Dunkin' cashiers becoming pharmacy techs. Thinking in skills and desired traits frees businesses and workers from obsessing over whether a job stays or goes. 'The winners in this AI era will be those who stop managing roles and start orchestrating skills,' said Joanna Riley, CEO of Censia. Make work suck less. AI can automate or reduce tedious job aspects, helping less experienced employees quickly take on more complex, meaningful, and better-paid work. 'The new versions of these jobs are just better — more interesting, creative, and human-oriented,' said Brandon Sammut, chief people officer at AI company Zapier. While there's a risk that these more meaningful and productive jobs will mean fewer jobs overall, the immediate opportunity is clear: Deploy AI tools that make jobs more enjoyable (less rote, routine work) and purposeful. Help people do harder stuff. Automating simpler tasks means work becomes more rewarding and challenging. Companies must help employees elevate their skill levels and continuously push into domains where humans have a unique advantage. 'We're headed toward a world where only high-skill, high-complexity work survives,' David Blake, CEO of Degreed, noted. 'The real challenge is helping people make that leap.' Durable human skills — like creativity, problem-solving, empathy, and mathematical reasoning — will remain essential. Support employees to use AI in daily work. It's management malpractice not to. 'It is management malpractice not to create an environment where employees are using AI in concrete ways in their daily work,' Sammut said. Knowing how to use AI is fast becoming table stakes to stay employable. The people most at risk of being replaced aren't being replaced by AI — they're being replaced by other humans who know how to use it. Helping employees integrate AI into their daily workflow doesn't just boost productivity — it also strengthens your employer value proposition. People are drawn to workplaces where they remain relevant, are continually challenged, and engage in meaningful, purposeful work. Waiting for certainty isn't an option. Leaders must act now, building strategies resilient to multiple outcomes. If you want to future-proof your business, start by future-proofing your people. Follow Allison Salisbury's writing and insights on her LinkedIn. Allison Salisbury is CEO of Humanist Venture Studio, which is supported by Stand Together, an organization that partners with changemakers who are tackling the root causes of America's biggest problems. Learn more about Stand Together's efforts to transform the future of work and explore ways you can partner with us.
Forbes
a day ago
- Forbes
Rigetti Computing Stock Down 37% As Nvidia CEO Sees ‘Inflection Point'
Nvidia CEO says quantum computing is reaching 'an inflection point' and Rigetti Computing stock is up slightly — but down 37% in 2025. Should you buy? Nvidia CEO Jensen Huang is now a quantum computing booster after a March conversion. Actively traded Rigetti Computing enjoyed a slight boost to its stock price which is down 37% in 2025 — after a 970% increase since June 2024. The company's recent earnings report does not support a bull case for Rigetti stock. Yet the longer-term opportunity for quantum computing could be significant, according to my June 10 interview with MIT expert Jonathan Ruane. Quantum computing — which uses super-cooled metal to power more calculations per circuit than classical computing can perform — is capturing the attention of investors. How so? Publicly-traded quantum computing stocks soared Wednesday before reversing direction later in the day. For example, according to Yahoo! Finance The reason? 'Quantum computing is reaching an inflection point,' Nvidia CEO Jensen Huang told a crowd at Nvidia GTC Paris Wednesday, reported Yahoo! Finance. 'We are within reach' of using quantum computers for 'areas that can solve some interesting problems in the coming years,' Huang added. This brings up an important observation about the industry: when Jensen Huang talks about quantum computing, people listen. That's my modern-day twist on the old catchphrase — 'When E.F. Hutton talks, people listen' — from a TV commercial for the late brokerage firm. Rigetti is among the most heavily traded public companies on the stock exchange — but should it be? The arguments opposing investor interest seem compelling: the company's stock has fallen 37% so far in 2025 and its first quarter 2025 financial results. Yet if prognosticators such as MIT Lecturer John Ruane are correct, the future of quantum computing could be very bright if certain problems are solved in the next five to ten years, he told me in a June 10 interview. Meanwhile Rigetti sees a bright future. 'We are four to five years from real commercial value of quantum computing,' Rigetti CEO Subodh Kulkarni told Asking For A Trend. 'That's where the market is really supposed to grow and be large enough where things like sales and EPS start becoming much more critical at that point. At this point, it's all about technology development and how we are getting the milestones done so that we enable this large $100 plus billion market in the future,' he added. Huang's views on QC have evolved significantly this year. He went from saying in January QC could be 15 to 30 years away from offering practical solutions to business problems, noted Investopedia, to seeing QC at an inflection point on June 11. During the interim, in March Huang announced a bet on a Boston-area research center to help Nvidia overcome the impediments to realizing the technology's potential, I wrote in my March Forbes post. At the moment, QC is based on fragile hardware and error rates that are unacceptably high for most practical applications. Classical computers rely on linear algebra, noted Yahoo! Finance. QCs, however, have far more processing capability because they apply quantum mechanics and advanced mathematics — making them amenable to solving problems in 'cybersecurity, cryptography, and chemistry,' reported Yahoo! Finance. Perhaps Huang was motivated to make his Wednesday remarks by IBM — which on June 10 announced plans to launch in 2029 a large-scale QC capable of 'operating without errors,' according to IEEE Spectrum. Google released a quantum computing chip called Willow in December, saying the technology "paves the way to a useful, large-scale quantum computer," noted Yahoo! Finance. While Amazon and Microsoft announced two quantum chips in February. Compared to these recent announcements, Huang's remarks were relatively content free. Meanwhile, Rigetti seems to have relatively little capital compared to these tech giants. What's more, the company says it is years away from being little more than a technology researcher hoping for a breakthrough from which the company can generate revenues. Rigetti's first quarter 2025 results were disappointing. The company's $1.5 million in revenue for the quarter fell more than 42% short of expectations while earnings per share of 13 cents missed by 84%, according to Google Finance. Rigetti net profit was less then met the eye. The company reported a $42.6 million profit in the first quarter of 2025, 'driven mainly by non-cash gains tied to financial instruments,' noted Quantum Insider. As Rigetti ramped up utility-scale quantum systems, the company revealed a $21.6 million operating loss. One thing Rigetti has going for it is the 970% increase in the company's stock price in the last year. The company has wisely capitalized on that increase to sell $350 million in common stock to raise its cash balance to about $570 million, according to Perhaps this cash will shore up the company's finances over the next four or five years as a hoped-for increase in revenue ultimately emerges. 'Quantum computing breakthroughs in science and engineering will make the tool more valuable for business. Better error correction algorithms will make QC less expensive and more consistent. Better performance will increase QC's commercial viability. Google's Willow chip showed big improvements in error correction,' Ruane told me in a June 10 interview. 'Engineering larger systems — including the chip and the surrounding infrastructure — that work together can improve QC's value to business. Google and IBM will operate across a rage of innovations to build powerful hardware platforms. Software startups will develop technology which venture capitalists will find amenable.' Ecosystems will be essential if the U.S. wants to lead the world in QC. 'Universities like MIT and Harvard will collaborate with large companies and startups to build ecosystems. Nvidia's Boston-based Accelerated Quantum Research Center is a good example. In May 2025, a new act was proposed — the quantum sandbox — aims to help create QC ecosystems,' he said. Problems with hardware and algorithms will need to be solved before achieving quantum advantage — where quantum surpasses classical computing. 'QC has the potential to make pharmaceutical development more efficient. This is not certain. There is a need to focus on hardware and algorithms. MIT and Harvard have algorithm developers like Aram Harrow,' he explained. It remains to be seen which problems will be more amenable to QC than classical computing. 'It is unclear whether QC will help make large language models easier to train and operate. Who could know what AI would end up doing?' 'There will never be a day when every person can have their own quantum computer — due to the physics. The use cases of QC are also unclear. QC may be operating more like supercomputers in hybrid clouds with classical computers. Everybody could end up using QC through apps powered by the hybrid cloud.' Business leaders will need new skills to take advantage of QC. Executives must 'learn how to reimagine solutions to previously unsolvable business problems using QC. This could be in areas such as logistics where a company is trying to optimize its supply chain.' If quantum advantage is near, will Rigetti be able to capitalize on it — or will IBM, Nvidia, Amazon, Google, and Microsoft get the lion's share?
CNBC
2 days ago
- CNBC
Trade tensions aren't stopping Chinese companies from pushing into the U.S.
BEIJING — Chinese companies are so intent on global expansion that even the biggest stock offering to date on Shanghai's tech-heavy STAR board counts the U.S. as one of its biggest markets, on par with China. Shenzhen-based camera company Insta360, a rival to GoPro, raised 1.938 billion yuan ($270 million) in a Shanghai listing Wednesday under the name Arashi Vision. Shares soared by 274%, giving the company a market value of 71 billion yuan ($9.88 billion). The United States, Europe and mainland China each accounted for just over 23% of revenue last year, according to Insta360, whose 360-degree cameras officially started Apple Store sales in 2018. The company sells a variety of cameras — priced at several hundred dollars — coupled with video-editing software. Co-founder Max Richter said in an interview Tuesday that he expects U.S. demand to remain strong and dismissed concerns about geopolitical risks. "We are staying ahead just by investing into user-centric research and development, and monitoring market trends that ultimately meet the consumer['s] needs," he told CNBC ahead of the STAR board listing. China launched the Shanghai STAR Market in July 2019 just months after Chinese President Xi Jinping announced plans for the board. The Nasdaq-style tech board was established to support high-growth tech companies while raising requirements for the investor base to limit speculative activity. In 2019, only 12% of companies on the STAR board said at least half of their revenue came from outside China, according to CNBC analysis of data accessed via Wind Information. In 2024, with hundreds more companies listed, that share had climbed to more than 14%, the data showed. "We are just seeing the tip of the iceberg. More and more capable Chinese firms are going global," said King Leung, global head of financial services, fintech and sustainability at InvestHK. Leung pointed to the growing global business of Chinese companies such as battery giant CATL, which listed in Hong Kong last month. "There are a lot of more tier-two and tier-three companies that are equally capable," he said. InvestHK is a Hong Kong government department that promotes investment in the region. It has organized trips to help connect mainland Chinese businesses with overseas opportunities, including one to the Middle East last month. Roborock, a robotic vacuum cleaner company also listed on the STAR board, announced this month it plans to list in Hong Kong. More than half of the company's revenue last year came from overseas markets. At the Consumer Electronics Show in Las Vegas this year, Roborock showed off a vacuum with a robotic arm for automatically removing obstacles while cleaning floors. The device was subsequently launched in the U.S. for $2,600. Other consumer-focused Chinese companies also remain unfazed by heighted tensions between China and the U.S. In November, Chinese home appliance company Hisense said it aimed to become the top seller of television sets in the U.S. in two years. And last month, China-based Bc Babycare announced its official expansion into the U.S. and touted its global supply chain as a way to offset tariff risks. Chinese companies have been pushing overseas in the last several years, partly because growth at home has slowed. Consumer demand has remained lackluster since the Covid-19 pandemic. But the expansion trend is now evolving into a third stage in which the businesses look to build international brands on their own with offices in different regions hiring local employees, said Charlie Chen, managing director and head of Asia research at China Renaissance Securities. He said that's a change from the earliest years when Chinese companies primarily manufactured products for foreign brands to sell, and a subsequent phase in which Chinese companies had joint ventures with foreign companies. Insta360 primarily manufactures out of Shenzhen, but has offices in Berlin, Tokyo and Los Angeles, Richter said. He said the Los Angeles office focuses on services and marketing — the company held its first big offline product launch in New York's Grand Central Terminal in April. Chen also expects the next phase of Chinese companies going global will sell different kinds of products. He pointed out that those that had gone global primarily sold home appliances and electronics, but are now likely to expand significantly into toys. Already, Beijing-based Pop Mart has become a global toy player, with its Labubu figurine series gaining popularity worldwide. Pop Mart's total sales, primarily domestic, were 4.49 billion yuan in 2021. In 2024, overseas sales alone surpassed that to hit 5.1 billion yuan, up 373% from a year ago, while mainland China sales climbed to 7.97 billion yuan. "It established another Pop Mart versus domestic sales in 2021," said Chris Gao, head of China discretionary consumer at CLSA. The Hong Kong-listed retailer doesn't publicly share much about its global store expansion plans or existing locations, but an independent blogger compiled a list of at least 17 U.S. store locations as of mid-May, most of which opened in the last two years. The toy company has been "very good" at developing or acquiring the rights to characters, Gao said. She expects its global growth to continue as Pop Mart plans to open more stores worldwide, and as consumers turn more to such character-driven products during times of stress and macroeconomic uncertainty.
